SYSTEM ONE PLAYGROUND
Jev answers typed questions about evidence you supply: a choice among your options, a yes or no, or a score on a written rubric. Each experiment below lets you edit the evidence, run a recorded example and read what came back, down to the exact request.
Recorded examples work here without an account. To run your own edits, use System One Cloud with its free daily allowance or your own engine.
6 experiments
Code checks contrast first. Jev compares the brief with the palette descriptions and can say that none fits.
Exact filters run in code first. Jev picks from the shortlist, checks each item's fit on its own and can report no match.
Send a batch of support reports once. A Choice, a yes/no question and a Score each read the same records.
Jev points to the passage that answers a question, checks whether any source answers it and flags sources that disagree.
Code reads the exit code and counts. Jev classifies the failure and suggests one of five next steps. The agent decides.
Two anchored Score questions rate clarity and tone. Watch the markers move as the draft improves.
No experiment matches. Try fewer words or another category.
Community and official demos worth studying. Each link goes to the author's own project. Several have no license file, so their code and assets stay with their authors; we link rather than copy.
A typed phrase re-themes the page. Code turns Choice and Score answers into colors and enforces contrast.
One Choice ranks commands, shown beside a fuzzy-match baseline so you can compare them.
Jev picks a bounded musical plan; code renders the sheet, audio and MIDI from that plan.
Editable A/B cases declare one input change before two runs. Recorded and live modes stay distinct.
Keyword results arrive first, then Jev's order slides in, with a yes/no check for whether any page answers.
Each run shows the request, the response and a separate box where your code decides.
Judgments update after typing pauses and a row lights up only when its answer meaningfully moves.
A terminal tone meter that ignores stale answers, and catalog choices with explicit no-match options.
One yes/no question per word, highlighted in three probability bands. The examples are in Japanese.
Test and build triage that checks whether an error exists before naming a root cause, with a receipt per call.
Compares an agent with and without a yes/no risk gate on proposed tool calls, including a planted log instruction.
Server time and end-to-end time shown side by side, with links that reopen a scenario.
Questions asked before anyone knows they matter, with code discarding the irrelevant answers.
Reviewed September 24, 2026. Speed, cost and accuracy figures that other projects report describe their own tests, not System One results.
Open the hosted Playground, sign in by email, and use your shared daily Jev allowance for live runs.
Add the System One MCP to Claude Code, Codex, Cursor or ChatGPT. Find your setup.
Recipes run with sysone_run; custom questions use sysone_decide. Each experiment shows the exact request to copy. Read the documentation.